AI SYSTEMS ENGINEERING

FAQ #

• What’s the “production AI wall”?

  PoCs work in controlled demos. Production is messier: multi-column PDFs, footnotes, niche vocabulary, 7-step workflows, and failing tool calls. The wall isn’t the model — it’s the engineering around it: agent state management, retrieval architecture, and fine-tuning. We build through it.
• When does RAG fail and how do you fix it?

  RAG usually fails at retrieval, not generation. Naive chunking splits context, cosine similarity returns adjacent but wrong chunks, and single-vector search misses exact matches. We fix it with hybrid retrieval, reranking, document-specific chunking, and retrieval evals before deployment.
• Is n8n / Make.com / Zapier involved in your work?

  Sometimes n8n is the right data pipe: webhooks, calendar triggers, moving records. We use reliable tools there and don’t reinvent them. But n8n isn’t the AI layer — complex agents, RAG, and fine-tuning need state, retrieval, and model engineering, and that’s where we fit.
• Do you work with our existing ML team?

  Yes — and we prefer it. Your ML team is strong on model training and research; we fill the production engineering layer: agent state management, retrieval architecture, inference optimization, and deployment infrastructure. We augment your team, not replace it.
• What if we just need a simple chatbot?

  Off-the-shelf RAG products like Cohere, Vertex AI Search, and OpenAI Assistants handle standard chatbot use cases well. If a managed template fits, we’ll say so in the CTO Consultation. We’re the right fit when you need agent architecture, custom retrieval, or domain fine-tuning.
• We built a PoC that hallucinates — can you fix it?

  Usually, yes. RAG hallucination is an engineering problem with diagnostics: wrong chunks, missing reranker, poor chunk strategy, or model output beyond context. We identify the failing layer and only rebuild if the architecture is fundamentally wrong — clarified in Discovery.
• Do you do AI strategy or just build?

  We build. The CTO Consultation maps your failure mode and scopes the architecture. The Discovery Phase validates the approach and produces the architecture spec. The Pilot Phase delivers working code. No strategy workshops, no transformation roadmaps, no AI readiness assessments.
• Do you sign NDAs?

  Yes. We sign NDAs before any technical discussion begins, on request.
• Who owns the code after delivery?

  You do. IP, source code, architecture docs, and deployment runbooks are fully assigned on completion — no vendor lock-in, no proprietary platform dependency.
• Do you work with formal contracts?

  Yes. Signed contracts with fixed scope per phase. You’re contracting with a registered entity, not an individual.